Compiling array computations for the Fresh Breeze Parallel Processor

Abstract

Fresh Breeze is a highly parallel architecture currently under development, which strives to provide high performance scientific computing with simple programmability. The architecture provides for multithreaded determinate execution with a write-once shared memory system. In particular, Fresh Breeze data structures must be constructed from directed acyclic graphs of immutable fixed-size chunks of memory, rather than laid out in a mutable linear memory. While this model is well suited for executing functional programs, the goal of this thesis is to see if conventional programs can be efficiently compiled for this novel memory system and parallelization model, focusing specifically on array-based linear algebra computations. We compile a subset of Java, targeting the Fresh Breeze instruction set. The compiler, using a static data-flow graph intermediate representation, performs analysis and transformations which reduce communication with the shared memory and identify opportunities for parallelization.